A COMPUTATIONALLY EFFICIENT COMPARISON OF ANN, CGPANN, AND CGP

Abstract

This paper presents a controlled empirical comparison of three evolutionary controllers for the CartPole-v1 control problem: a conventional artificial neural network (ANN), a Cartesian genetic programming encoded neural network (CGPANN), and a classical or full Cartesian genetic programming (CGP) controller. To isolate the effect of representation, all methods use exactly 10 computational nodes and are optimized with the same (1+5) evolutionary strategy, 500 generations, 0.10 mutation probability, and random seed 42. The study evaluates not only task fitness but also representation size and approximate computational cost. Under the reported single-seed experiment, the ANN reaches a best fitness of 14 steps, CGPANN reaches 10 steps, and Full CGP reaches 8 steps. In contrast, inference cost decreases from 200 FLOPs/action for ANN to 142 FLOPs/action for CGPANN and 55 FLOPs/action for Full CGP. Total evolutionary training cost falls from 7.0026 million FLOPs for ANN to 3.5514 million for CGPANN and 1.1004 million for Full CGP. These results show a clear trade-off. The more compact representations in this implementation reduce arithmetic cost substantially, but they do not achieve the same CartPole fitness as the dense ANN under the tested evolutionary budget. The findings motivate future multi-seed experiments and improved CGP selection mechanisms for testing whether the observed efficiency can be retained while increasing control performance.